A scientific research achievement intelligent examination method and system for fund project progress and completion report

Abstract

The application discloses a kind of scientific research achievements intelligent examination method and system for fund project progress and conclusion report, it is related to scientific research project management field, the method includes: report file is parsed and cleaned, extracts the project identification information of title page, locates achievement chapter and is divided into independent item, extracts achievement information.Based on the composite cache key formed by project approval number, achievement category and achievement name carries out cache determination, when not hit, according to the dynamic routing of achievement type and language attribute to the adapted external database for retrieval according to achievement type and language attribute, with the aid of the semantic similarity algorithm based on Token set and double-element joint determination realizes noise-tolerant type achievement matching.The heterogeneous metadata returned by retrieval is mapped to uniform verification field, fusion rule verification is executed, and structured examination conclusion is output and classified archiving.The application can effectively improve the accuracy of achievement item identification, multi-source retrieval hit rate and batch verification efficiency, and realize traceable automatic examination.

Description

Technical Field

[0001] This invention relates to the field of scientific research project management and intelligent information processing technology, and in particular to a method and system for intelligent review of scientific research achievements for funded project progress and final reports. Background Technology

[0002] In the actual management of National Natural Science Foundation of China (NSFC) projects, various problems are frequently discovered during the review of research project results. For example, papers may not properly indicate the NSFC grant number; the publication date or patent application / authorization date may not match the project execution period; and there may be inconsistencies between the patent holder and the affiliated institution. Due to the large number of project progress reports and final reports submitted at highly concentrated times, and the significant differences in retrieval methods and presentation formats of results across different academic databases, relying solely on manual review by the affiliated institution's research management department would result in excessive workload, low processing efficiency, and difficulty in creating a complete review record. Failure to promptly identify and urge project leaders to rectify these issues will directly impact the quality of project completion and the management integrity of the affiliated institution.

[0003] To address the aforementioned needs, existing technologies primarily offer three types of solutions. The first is manual verification, which relies on the experience of reviewers to meticulously check each item in the report. This method is slow, inconsistent in its review standards among different personnel, and often fails to achieve full coverage when dealing with a large number of items. The second is general document recognition technology, which uses optical character recognition engines to extract text from report files. However, this method only obtains the raw text and cannot present the items in a structured form, making it impossible to directly enter a rule-based automated verification process. The third is single-database retrieval, which retrieves information from a single academic database. However, different databases have significantly different field definitions and formats, and lack a mechanism for integrating and judging key information such as project approval number, implementation period, and affiliated institution, leading to insufficient accuracy in verification conclusions.

[0004] At the text parsing level, the portable document format files of progress reports and final reports are weakly structured long texts with mixed distribution of numbering, line breaks, and parenthetical notes, resulting in poor stability in identifying achievement items. In scenarios with mixed Chinese and English text, the boundaries between paper titles and author fields are unclear, easily leading to retrieval errors. At the database integration level, the field structures returned by different databases are heterogeneous, and fields such as publication date, grant number, and patent holder lack a unified mapping standard, making it difficult to guarantee the stability of rule-based judgment. At the batch processing level, existing solutions lack caching and reuse mechanisms and archiving designs, resulting in high time and data query costs due to repeated retrieval when dealing with large batches of reports. Summary of the Invention

[0005] This invention overcomes the shortcomings of existing technologies and provides an intelligent review method and system for scientific research achievements, focusing on the progress and completion reports of funded projects. It employs a constraint-driven semantic-aware localization method for weakly structured long text achievements, stably identifying chapter boundaries even in highly irregular document structures and eliminating interference from text citations. A title boundary inference algorithm based on academic literature citation formats is used, implementing dual-path extraction for mixed Chinese and English entries to accurately extract search key fields such as achievement name, author, and year. A language-aware dynamic routing strategy for search sources and a heterogeneous achievement type matching strategy are constructed, dynamically selecting the most suitable database and matching algorithm based on achievement type and language attributes, and mapping heterogeneous metadata to unified verification fields. Furthermore, a token-based semantic similarity algorithm and a two-factor joint judgment mechanism are used to achieve noise-tolerant achievement matching. Simultaneously, a composite cache key manages the verification status across batches, forming traceable and interpretable compliance verification conclusions.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Firstly, the present invention provides an intelligent review method for research achievements related to the progress and completion reports of funded projects, comprising the following steps:

[0007] S1. Receive and parse the progress report or completion report of the funded project, perform text standardization and cleaning, and extract the project identification information from the first page of the report; the project identification information includes at least the project approval number, the affiliated institution, and the project execution period;

[0008] S2. Locate the achievement section in the cleaned report text, divide the achievement section into independent achievement items, and extract information from each achievement item in a structured manner to obtain achievement information that includes at least the achievement name, achievement category, author, and publication year.

[0009] S3. Based on the composite cache key consisting of the project approval number, achievement category, and achievement name of the current achievement item, query and verify the cache record; if a match is found, mark it as verified and skip the subsequent search steps; if no match is found, proceed to the multi-source search step.

[0010] S4. Based on the category and language attributes of the results, dynamically determine and call at least one external academic database for retrieval, and use a semantic similarity algorithm based on token sets to match the retrieval results with the results information to obtain candidate matching results;

[0011] S5. Obtain the metadata of the candidate matching results, map it to the unified verification field, and perform fusion rule verification with the project identification information and result information extracted from the report homepage to generate item-level verification conclusions; the fusion rule verification includes at least the verification of funding agency and approval number, result time window verification, and patent holder consistency verification.

[0012] S6. Output the full retrieval process and verification conclusions in a structured manner, and append the information of the verified result items to the verification cache record.

[0013] In a preferred embodiment of the present invention, in step S2, the positioning result section specifically includes:

[0014] A constraint-driven semantic perception and localization method for weakly structured long text outputs is adopted. The expected order of occurrence of output chapter names is preset, and all candidate positions of each chapter name are retrieved in the full text.

[0015] The candidate set is filtered by the constraint that the index values ​​of each candidate position are strictly monotonically increasing, and the set of chapters that satisfy the desired order and whose position values ​​are continuously increasing is retained.

[0016] For frequently occurring keywords, the adjacent suffix pattern is detected through a contextual sliding window, and candidate positions that match the exclusion pattern are filtered during the retrieval stage.

[0017] In a preferred embodiment of the present invention, in step S2, during the structured extraction of information, for English achievement entries, a result title boundary inference algorithm oriented towards academic literature citation formats is specifically adopted:

[0018] Filter out Chinese characters in the entries to select a set of valid English tokens;

[0019] The trailing anchor tokens are determined in four priority levels: the first priority is the journal or conference name token containing compound academic identifier phrases; the second priority is the token containing conference identifiers and also containing year numbers within a preset year range; the third priority is the year token consisting of four digits; and the fourth priority is the second to last item in the valid token set as the fallback anchor.

[0020] Using the determined anchor token as the front boundary delimiter, the first segment of the article is truncated and the last semicolon-separated unit is extracted as the first title candidate; at the same time, the token with the most words, after excluding tokens already identified as journal names or conference names, is used as the second fallback candidate; the two together form a set of multiple candidate titles.

[0021] In a preferred embodiment of the present invention, in step S4, dynamically determining the calling order specifically involves:

[0022] A two-dimensional routing matrix is ​​constructed with output type and language attribute as the dimensions. The determination of the language attribute of the output adopts a two-stage mechanism: first, the proportion of Chinese characters is counted. When it exceeds the preset threshold, it is determined to be Chinese. Otherwise, a statistical language model is used for classification.

[0023] For journal articles or conference papers in English, English databases are used first, followed by Chinese databases; for journal articles or conference papers in Chinese, Chinese databases are used first.

[0024] For patent-related achievements, the patent database is specifically accessed, and a precise matching strategy based on the patent application number is used for matching.

[0025] In a preferred embodiment of the present invention, in step S4, the semantic similarity algorithm specifically comprises:

[0026] Set the first threshold. Second threshold and author auxiliary threshold ,in ;

[0027] For each candidate record in the search results list, calculate the semantic similarity between its result name and the target result name based on the token set. And the similarity between the author list and the target author. ;

[0028] when When it is directly judged as a hit; when and When the time is right, it is considered a hit;

[0029] Take all hit candidates The largest one is used as the final matching result when If a perfect score is obtained, the subsequent search will be terminated immediately.

[0030] In a preferred embodiment of the present invention, in step S5, the result time window verification specifically adopts a differentiated domain semantic time window expansion model:

[0031] The total start date of the time window is September 1st of the calendar year preceding the start date of the project execution period, and the total end date of the time window is the calendar year following the end date of the project execution period.

[0032] For academic papers, the starting point of the time window will be adjusted to January 1st of the year in which the implementation period begins;

[0033] For patent-related achievements, the patent application date or the date of authorization announcement is used as the time determination field.

[0034] In a preferred embodiment of the present invention, in step S5, the verification of the funding agency and approval number further includes:

[0035] When the funding category of a project belongs to a type with a sub-project hierarchical structure, the prefix field of the parent project approval number is taken and matched with the sub-project approval number by truncation and inclusion.

[0036] In a preferred embodiment of the present invention, in step S5, the patent holder consistency verification specifically employs a keyword inclusion matching strategy:

[0037] The verification is considered successful when the patentee field returned by the search contains any of the following forms: the full name, abbreviation, English name of the affiliated unit, or the name of its subordinate second-level unit.

[0038] In a preferred embodiment of the present invention, the method further includes: managing the processing results of each report hierarchically based on the document-level verification pass rate, and copying all related products generated by the same report to the corresponding level directory to form a visual hierarchical archiving structure.

[0039] Secondly, this invention provides an intelligent research achievement review system for an intelligent review method of research achievement progress and completion reports for funded projects, comprising:

[0040] The results extraction module is used to receive and parse the progress report or completion report of the funded project, perform text standardization and cleaning, extract project identification information from the first page of the report, locate the results chapter in the cleaned report text and divide it into independent results items, and perform information structure extraction on each results item to obtain results information that includes at least the results name, results category, author and publication year.

[0041] The verification cache management module is used to query verification cache records based on a composite cache key consisting of the project approval number, achievement category, and achievement name of the current achievement item. If a match is found, it is directly marked as verified and subsequent searches are skipped. If no match is found, the multi-source search module is triggered. The module is also used to append the information of verified achievement items to the verification cache record.

[0042] The multi-source retrieval module is used to dynamically determine and call at least one external academic database for retrieval based on the category and language attributes of the results, and to use a semantic similarity algorithm based on token sets to match the retrieval results with the results information to obtain candidate matching results.

[0043] The consistency verification module is used to obtain the metadata of candidate matching results, map it to a unified verification field, and perform fusion rule verification with the project identification information and result information extracted from the report homepage to generate item-level verification conclusions; the fusion rule verification includes at least the verification of funding agency and approval number, result time window verification, and patent holder consistency verification.

[0044] The results output module is used to output the full retrieval process and verification conclusions in a structured manner.

[0045] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0046] (1) This invention provides an intelligent review method and system for research achievements, including progress reports and final reports of funded projects. It employs a constraint-driven semantic perception and positioning method for weakly structured long text achievement chapters. This method uses a preset expected sequence of chapter names and a strictly monotonically increasing index value for each candidate position as a constraint for filtering. Simultaneously, it excludes non-chapter start features using a contextual sliding window for high-frequency keywords. This allows the system to distinguish between false positioning caused by text citations, annotations, or tables and true chapter titles, even in cases of highly irregular document structures. This improves the accuracy of chapter boundary positioning, and the text range received in subsequent item segmentation and information extraction steps is also more precise. Compared to existing solutions that rely solely on the first occurrence of keywords or fixed offsets, this invention effectively avoids chain errors caused by mispositioning of chapters, thus ensuring the reliability of the entire review process.

[0047] (2) This invention employs a title boundary inference algorithm oriented towards academic literature citation formats. Its multi-level priority and dual-candidate design logic enables the system to achieve robust title boundary inference for non-standardized academic citation formats without relying on predefined dictionaries or machine learning models. This effectively ensures the recall rate of the results titles and significantly reduces the risk of missed detections in subsequent searches due to title offsets. Compared to the problem of boundary recognition failure in general natural language processing tools when dealing with academic citation texts with mixed multi-field structures, this invention can still maintain stable extraction performance in scenarios with highly heterogeneous citation formats.

[0048] (3) This invention employs a language-aware dynamic routing and heterogeneous result type matching strategy. A two-dimensional routing matrix is ​​constructed using result type and language attributes to dynamically determine the order of database calls and matching strategies. The combination of classification routing and type-aware matching ensures that all types of results can hit the most suitable data source and matching algorithm, thereby effectively reducing invalid searches and simultaneously improving the overall search hit rate and verification accuracy. Compared to technologies that uniformly use fuzzy title matching for all result types or rely on a single database, this invention eliminates the risk of misjudgment caused by inconsistent language styles in patent names and achieves differentiated and efficient retrieval for Chinese and English results.

[0049] (4) This invention employs a noise-tolerant two-factor semantic matching mechanism. Its design for joint confidence judgment of two factors enables the system to effectively eliminate interference from format noise such as word order misalignment, punctuation residue, line break embedding, and mixed simplified and traditional Chinese characters. Furthermore, while tolerating slight title distortions, the introduction of author information as an auxiliary constraint keeps the misjudgment rate caused by homonymous variant texts within a reasonable range. Compared to existing solutions that rely on precise string matching or a single similarity threshold, this invention exhibits stronger matching robustness, and its short-circuit termination strategy further reduces the cost of invalid queries in batch processing scenarios.

[0050] (5) This invention constructs a domain-adaptive multi-dimensional compliance verification rule system, encoding domain knowledge into configurable verification rule designs. This allows time-based judgments to accurately cover submitted manuscripts in the pre-research stage and results that have not yet been formally published due to publication cycle delays in the final stage. It also adapts to the differences in approval number formats under different funding levels, thereby effectively reducing misjudgments and omissions caused by publication cycle characteristics and differences in project funding levels. Compared to general solutions that rely solely on the start and end dates of execution as a unified judgment boundary or perform precise matching of all fields, the consistency between the batch verification conclusions and the manual review conclusions of this invention is higher.

[0051] (6) This invention establishes an incremental idempotent verification status management mechanism based on composite keys. Its idempotent design ensures that the system does not need to restart from scratch after a restart in scenarios such as network anomalies, database response timeouts, or batch interruptions. Only the incomplete parts are processed, thereby significantly reducing the cost of repeated retrieval across batches and substantially improving the processing efficiency in large-scale review scenarios. At the same time, hierarchical archiving based on file-level verification pass rates allows managers to directly locate high-risk reports through the directory structure without having to open and review them one by one. Compared with solutions that lack cache reuse and hierarchical archiving mechanisms, this invention can significantly reduce the number of external database queries in actual batch reviews and achieve traceable audit output. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings;

[0053] Figure 1 is a flowchart of the intelligent review method for the scientific research achievements of the present invention;

[0054] Figure 2 is an architecture diagram of the intelligent review system for the scientific research achievements of the present invention. Specific embodiments

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0056] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention, but the present invention may be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0057] Embodiment 1

[0058] This embodiment provides an intelligent review method for scientific research achievements for the progress and final report of fund projects, aiming to solve the problems of low review efficiency, poor accuracy, and difficulty in traceability caused by weak report text structure, heterogeneous fields, and inconsistent multi-source data caliber when the scientific research project management department batch-processes project progress reports and final reports.

[0059] Refer Figure 1 , which shows the overall flowchart of the method described in this embodiment. The execution subject of this method is a computer device configured with at least one processor and one memory, and instructions executable by the processor are stored in the memory to implement the following steps:

[0060] S1. Receive and parse the progress report or final report file of the fund project, perform normalization cleaning on the text, and extract project identification information from the report front page.

[0061] In this embodiment, one or more progress reports or final report files of fund projects uploaded by the user through the interaction interface are received, and the files are documents in portable document format or OFFICE OPEN XML format.

[0062] Specifically, in document parsing and content extraction, a stable version of the document parsing library is called to losslessly convert the binary format report file into a more structured document object model, thereby obtaining the character stream of the full report and its rough layout position information.

[0063] Because the original parsed text often contains noisy characters unrelated to the content, such as page numbers, headers and footers, page breaks, and continuous spaces, a regular expression matching and replacement strategy is used for batch cleaning. For example, patterns are defined to match single or multi-page page numbers, and a compiled finite automaton is used to quickly locate and delete these patterns throughout the entire text. Simultaneously, various whitespace characters are uniformly mapped to a single standard space character, forming a clean, continuous character sequence, providing an interference-free text foundation for locating subsequent chapters.

[0064] Furthermore, the cleaned text header, such as the first 5000 characters, is scanned, and a series of extractors based on context-free grammars are used to capture key item identification information fields. Specifically:

[0065] Project approval number extractor: Matches a specific pattern consisting of numbers and hyphens, such as [0-9]{7,8} or other prefixed approval numbers.

[0066] Supporting Institution Extractor: Captures the institution name after clearly identifying labels such as supporting institution and undertaking institution.

[0067] Execution Year Extractor: Captures the start and end years following the execution year label and converts them into a unified machine-readable format, recording the start year as... The cutoff year is denoted as The execution period is then converted into two absolute time boundaries for subsequent verification: the starting point of the verification time window. and the end of the verification time window . The calculation method is as follows September 1st of that year, The calculation method is as follows December 31st of the year. This calculation method takes into account the industry practice of submitting preliminary research papers before the formal implementation of funded projects.

[0068] S2. Locate the results section in the cleaned report text, divide the results section into independent results entries, and perform information structure extraction on each results entry to obtain results information that includes at least the results name, results category, author, and publication year.

[0069] Specifically, it is further subdivided into three sub-steps: result chapter positioning, item segmentation, and structured extraction.

[0070] In this embodiment, the result chapter location adopts a constraint-driven semantic-aware location method for weakly structured long text result chapters; a preset expected order sequence of result chapter names is maintained in memory, denoted as... Here, Sk represents a set of keywords for the title of a research paper's chapter. For example, in a National Natural Science Foundation of China report, S1 = {journal article, magazine article}, S2 = {conference paper, academic conference}, S3 = {patent, invention patent, utility model patent}, etc.

[0071] Among them, in the cleaned full text, for For each Sk, retrieve all possible starting position indices to form a candidate position set. For example, for the keyword "patent," the system will not only match independent patent chapter titles, but may also match pseudo-positions such as "this technology has been patented" in the text.

[0072] Continue performing constraint filtering and iterating through the Cartesian product. Search for all paths that satisfy the strictly monotonically increasing condition, i.e., find a set of index values. , making For candidate paths that meet this condition, only the first candidate position corresponding to each chapter name on that path is retained; the remaining positions are judged as false locations and are eliminated.

[0073] For keywords such as "patent" that appear frequently in text, this method actively excludes them during the retrieval stage. It uses a contextual sliding window to detect the immediate suffix patterns of the candidate keyword, such as defining a set of key symbols or keyword exclusion patterns. If any pattern in the exclusion pattern set is matched within 1-5 characters after the candidate keyword, that candidate position is marked as invalid during the generation stage and will not proceed. This sets up a collection, thereby reducing the generation of false candidates at the source.

[0074] Through a multi-level recovery chain consisting of candidate generation, monotonic filtering, and context exclusion, the true boundaries of the output chapters can be stably and accurately located.

[0075] Furthermore, the entry segmentation is as follows: after determining the starting position of the achievement chapter, the first-level number within the chapter, such as "1.", "2.", "[1]", "(1)", etc., is used as the separator. Regular expressions are used to segment the chapter text, breaking the continuous text into independent, unstructured original achievement strings. Each string corresponds to one achievement (a paper or a patent).

[0076] Furthermore, the structured extraction of results information specifically involves the system performing structured information extraction for each original result string to obtain structured result information, including result category, title, author, year, etc. This method uses different extraction paths for items with mixed Chinese and English text.

[0077] For entries identified as primarily in Chinese, the system uses the longest valid Chinese segment rule for parsing. The core of this rule is that in Chinese citation formats, information such as titles, authors, and journal names are typically separated by periods, commas, spaces, etc., and the title is usually the longest and most continuous Chinese string in the entry. The system calculates the length of each continuous Chinese string in the entry and extracts the longest string as the achievement title. Content before the achievement title is used as candidate authors, and content after the achievement title and before the year is used as candidate journal / conference names.

[0078] For research entries that are primarily in English, their citation formats are complex and varied. This method employs an innovative algorithm for inferring the boundaries of research titles based on academic literature citation formats.

[0079] Specifically, Chinese characters and non-alphanumeric characters are filtered from the original output string to select the English token sequence consisting of letters. The header's trailing anchor token is determined according to the following four priority levels:

[0080] The first priority is for compound academic identifier anchors: scan the English token sequence to see if it contains an entry from a predefined library of compound academic identifier phrases, which includes, but is not limited to, Journal of, Proceedings of, Symposiumon, Transactions on, Conference on, etc. Tokens that match are considered subsequent bound anchors.

[0081] The second priority is the conference identifier + year anchor: if the first priority is not found, then look for a token that meets both conditions: a) contains the conference identifier; b) is near the token, such as a previous or next token within a preset year range. The token is set as the backbound anchor for the four-digit year within the specified range.

[0082] The third priority is the pure year anchor: if the first two levels are not matched, then search for a four-digit number in the English token sequence that falls within the specified range. Set the token within the bounds as the trailing anchor.

[0083] The fourth priority is the fallback anchor: If the anchor cannot be determined in the above three levels, then the second to last token in the English token sequence, i.e., index N-1, is taken as the fallback anchor.

[0084] Using the trailing anchor point as the preceding delimiter, all tokens from the beginning of the English token sequence to the trailing anchor point are extracted to form a title candidate prefix. Within this prefix, the last semicolon ";" or "," is searched, and the content after that semicolon is extracted as the first title candidate. Simultaneously, all tokens identified as journal or conference names are excluded from the English token sequence. From the remaining tokens, the consecutive token string with the most words is selected as the second title candidate.

[0085] Understandably, by combining the first and second title candidates into a multi-candidate title set and prioritizing the former for subsequent retrieval, the robustness of English title extraction can be greatly improved without relying on pre-trained models through four-level priority and dual candidate protection.

[0086] S3. Based on the composite cache key consisting of the project approval number, achievement category, and achievement name of the current achievement entry, query and verify the cache record; if a match is found, mark it as verified and skip the subsequent search steps; if no match is found, proceed to the multi-source search step.

[0087] It should be noted that in batch processing scenarios, in order to avoid repeated retrieval and verification of the same result, an idempotent verification status management mechanism is introduced. This mechanism relies on a verification cache file stored in a local file system or a lightweight embedded database.

[0088] Specifically, composite cache keys The definition of is a string composed of three fields concatenated together:

[0089] .

[0090] in, The project approval number; Category of achievement; To extract the hash value of the Message Digest Algorithm Version 5 from the names of the cleaned results; This is used as a separator to distinguish different field values ​​(project approval number, result category, result name hash value). Using the hash value instead of the original name as part of the key can effectively avoid interference from special characters in the result name, such as newlines and quotation marks, on cached queries and improve query efficiency.

[0091] Furthermore, when a new outcome item is pending processing, its... The system then queries the cache file. If the query finds a match, it indicates that the result has been successfully verified in a previous batch of runs. The system directly reads the historical verification results of the result from the cache record and marks it as verified - cache hit. No further steps are required, thereby reducing the external database query pressure and shortening the overall processing time.

[0092] S4. Based on the category and language attributes of the results, dynamically determine and call at least one external academic database for retrieval, and use a semantic similarity algorithm based on token sets to match the retrieval results with the results information to obtain candidate matching results.

[0093] Specifically, for results that do not hit the cache, the process proceeds to the multi-source intelligent retrieval stage. The core of this stage is a language-aware dynamic routing of retrieval sources and a heterogeneous result type matching strategy.

[0094] The determination of the language attributes of the output adopts a two-stage mechanism:

[0095] Phase 1 is character ratio determination: count the number of Unicode Chinese characters in the entry. If the proportion of Chinese characters without spaces to the total number of characters exceeds a preset threshold (preferably 0.3), it is directly determined to be Chinese.

[0096] Phase two involves a statistical language model assessment: if the threshold in Phase one is not exceeded, a statistical language model based on character n-grams is activated. This model is pre-trained on Chinese and English corpora and can calculate the probability that a character sequence belongs to Chinese or English respectively. By comparing the probability values, accurate language classification is performed on short texts with ambiguous boundaries.

[0097] Furthermore, a two-dimensional routing matrix is ​​constructed with output category and language attribute as dimensions, and the following routing rules are defined:

[0098] For entries where the category of output is "paper" and the language is "English", the routing order is: Scopus database - CNKI - Wanfang Data. Scopus database, which is internationally recognized and has the most comprehensive coverage of English literature, is searched first. Chinese databases are only used as a supplement if no high-confidence match is found in Scopus, in order to prevent situations where there is a Chinese version of the output or both Chinese and English versions exist.

[0099] For entries with output category = paper and language = Chinese, the routing order is: CNKI - Wanfang Data.

[0100] For entries with the achievement category equal to patent, the routing strategy is completely different from that for papers. They will be directly routed to the CNKI patent database and Wanfang Data patent database, and processed by a unified patent search interface.

[0101] Furthermore, after retrieving a set of candidate results from an external database, this method employs a semantic similarity algorithm based on token sets for accurate matching. This algorithm can effectively resist interference such as word order misalignment and punctuation noise.

[0102] Specifically, let the title of the target result be string A. After standardized word segmentation and stop word filtering, it yields an independent set of tokens. Similarly, if the title obtained from the candidate results is B, then a Token set is obtained. .

[0103] Among them, semantic similarity is defined. The calculation formula is as follows:

[0104] .

[0105] in, The intersection of two sets is represented by the set that is simultaneously in the same place. and Tokens appearing in; and These represent the size of the two sets respectively; this formula is essentially a representation of cosine similarity in a set context, effectively mitigating the influence of token order.

[0106] Author similarity The calculation uses the Jaccard similarity coefficient:

[0107] .

[0108] in, and These are the sets of author surnames or full names for the target author list and the candidate author list, respectively.

[0109] It should be noted that the preset first threshold 0.85, second threshold 0.65 and author auxiliary threshold The value is 0.5. The matching logic is as follows:

[0110] like It is directly judged as a hit.

[0111] like and Similarly, it is also judged as a hit. This dual-factor joint judgment mechanism can tolerate slight deformation of the title, while using author information as a strong constraint, which can effectively reduce the misjudgment rate of variant texts with the same name.

[0112] After iterating through all recall candidates, the system selects the hit candidates. The one with the largest value is taken as the final match result, once found. For candidates that score 1.0, immediately interrupt subsequent database calls and comparisons of the remaining candidates to achieve short-circuit termination and save computing resources.

[0113] If no matching result is found in any of the configuration databases, the reason for the failure of this search will be recorded in the entry log. This entry will not be included in the verification cache in this batch, and the search will be re-initiated on the next run.

[0114] S5. Obtain the metadata of the candidate matching results, map it to the unified verification field, and perform fusion rule verification with the project identification information and result information extracted from the report homepage to generate item-level verification conclusions.

[0115] Metadata fields returned by different databases, such as date, funding information, and rights holders, have different formats. These are mapped to a set of unified verification fields defined internally by the system, namely uniform_publication_date, uniform_funding_info, and uniform_assignee, for verification using integrated rules. The core of this system is a domain-adaptive multidimensional compliance verification rule system.

[0116] In terms of time window calculation, unlike the general approach that uses the start and end years of execution as fixed boundaries, this method designs a differentiated time window extension model with domain semantics to accurately cover the situation of pre-research submission and publication delays.

[0117] Specifically, pre-calculate the global verification boundary:

[0118] Verification time window start point .

[0119] End of verification time window .

[0120] Based on this, the results of papers and patents are treated differently:

[0121] For academic papers, including journal articles and conference papers, the starting point of the time window has been adjusted to... Extract the publication year of the paper. If the date is accurate to the day from the metadata, use the specific date; otherwise, if the date is within the range of... If the time frame is within the specified range, the time verification will be successful. This adjustment can cover submissions that were completed but not yet published before January 1st of the first year of the implementation period.

[0122] For patent-related achievements, the patent application date or authorization announcement date is extracted according to priority and recorded as follows: ,like In If the time frame is within the specified interval, the time verification passes; a more lenient global time window is then used. This covers patent applications made one year before the start of the project execution period.

[0123] Furthermore, regarding the matching of project approval numbers, the funding information field uniform_funding_info in the results metadata is verified to confirm whether it contains the National Natural Science Foundation of China or its standard English name "National Natural Science Foundation of China".

[0124] For the approval number, this method supports sub-project scenarios. It retrieves the funding category field of the project from the information on the first page of the report. If it is marked as a major project or a type with a clear sub-project hierarchical structure, the truncation inclusion matching mode is activated.

[0125] Specifically, a prefix is ​​extracted from the parent project approval number, i.e., the approval number of the current report. For example, if the parent project approval number is "12345678", the first 6 digits "123456" are extracted as a truncated prefix. Then, the fund approval number marked in the result metadata is checked. If its string contains the truncated prefix, the approval number verification is passed, thereby avoiding misjudgment caused by the difference in the suffix of the factor project.

[0126] Furthermore, regarding the patentee, the `uniform_assignee` field is retrieved from the metadata of the achievement. Simultaneously, the full name of the affiliated institution is obtained from the information on the report's homepage, and its abbreviation, English name, and common names of subordinate units are generated using a pre-defined thesaurus. For example, Fudan University corresponds to Fudan University, Fudan University Affiliated Zhongshan Hospital, etc. The longest common substring comparison algorithm is used to determine if a match exists. If the length of the longest common substring of the `uniform_assignee` string exceeds a threshold with any of the above forms, the verification is considered successful, thus resolving the issue of inconsistent representations of institution names across different databases.

[0127] All the above verifications are performed on an item-by-item basis, and all verification results are encapsulated into an item-level verification conclusion object.

[0128] S6. Output the full retrieval process and verification conclusions in a structured manner, and append the information of the verified result items to the verification cache record.

[0129] In this embodiment, a comprehensive audit package is generated, which generates retrieval and verification log files and summary statistical table files for each report.

[0130] The retrieval and verification log file records the full lifecycle information of each deliverable entry in the report in a structured format, including: input deliverable information, a list of databases invoked, retrieval requests and responses for each database, and intermediate values ​​of semantic similarity calculations. and The final matching candidate results, and the detailed basis and conclusions for verifying each rule.

[0131] The summary statistics table file, in comma-separated value format, records the final judgment conclusion for each achievement item and the specific evidence of failure.

[0132] For all entries whose verification result is "passed", set their composite cache key. The current verification timestamp is appended to the persistent verification cache file, ensuring that the result will be skipped in the next batch processing run, thus achieving an incremental and idempotent verification process.

[0133] Example 2

[0134] like Figure 2 As shown, this embodiment provides an intelligent review system for research achievements related to the progress and final report of funded projects, including:

[0135] The results extraction module is used to receive and parse progress reports or completion reports of funded projects, perform text standardization and cleaning, extract project identification information from the first page of the report, locate the results chapters in the cleaned report text and divide them into independent results entries, and perform information structure extraction on each results entry to obtain results information that includes at least the results name, results category, author, and publication year.

[0136] The verification cache management module is used to query verification cache records based on a composite cache key consisting of the project approval number, achievement category, and achievement name of the current achievement item. If a match is found, the record is marked as verified and subsequent searches are skipped. If no match is found, the multi-source search module is triggered. The module is also used to append verified achievement item information to the verification cache record.

[0137] The multi-source retrieval module is used to dynamically determine and call at least one external academic database for retrieval based on the category and language attributes of the results, and uses a semantic similarity algorithm based on token sets to match the retrieval results with the results information to obtain candidate matching results.

[0138] The consistency verification module is used to obtain the metadata of candidate matching results, map it to the unified verification field, and perform fusion rule verification with the project identification information and result information extracted from the report homepage to generate item-level verification conclusions. The fusion rule verification includes at least the verification of funding institutions and approval numbers, result time windows, and patent holder consistency verification.

[0139] The results output module is used to output the full retrieval process and verification conclusions in a structured manner.

[0140] It should be noted that the intelligent scientific research results review system can implement the steps in the intelligent scientific research results review method in Embodiment 1 above, and can achieve the same technical effect. Refer to the description in the above embodiments, which will not be elaborated here.

[0141] The above description is based on the preferred embodiments of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0142] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for intelligent review of research achievements related to the progress and final report of funded projects, characterized in that, Includes the following steps: S1. Receive and parse the progress report or completion report of the funded project, perform text standardization and cleaning, and extract the project identification information from the first page of the report; the project identification information includes at least the project approval number, the affiliated institution, and the project execution period; S2. Locate the achievement chapters in the cleaned report text and divide the achievement chapters into independent achievement items. Extract information from each achievement item in a structured manner to obtain achievement information that includes at least the achievement name, achievement category, author, and publication year. Use a constraint-driven semantic perception localization method for weakly structured long text achievement chapters to locate the achievement chapters. The expected order of appearance of the achievement chapter names is preset, and the candidate set is filtered by the constraint that the index value of each candidate position is strictly monotonically increasing. S3. Based on the composite cache key consisting of the project approval number, achievement category, and achievement name of the current achievement item, query and verify the cache record; if a match is found, mark it as verified and skip the subsequent search steps; if no match is found, proceed to the multi-source search step. S4. Based on the category and language attribute of the achievement, dynamically determine and call at least one external academic database for retrieval; adopt a two-stage mechanism to determine the language attribute of the achievement: first, count the proportion of Chinese characters, and if it exceeds a preset threshold, it is determined to be Chinese; otherwise, a statistical language model is used for classification; and use a semantic similarity algorithm based on token set to match the retrieval results with the achievement information to obtain candidate matching results; the semantic similarity algorithm sets a first threshold, a second threshold, and an author auxiliary threshold, and adopts a two-factor joint judgment mechanism to achieve noise-tolerant achievement matching; S5. Obtain the metadata of the candidate matching results, map it to a unified verification field, and perform fusion rule verification with the project identification information and result information extracted from the report homepage to generate an item-level verification conclusion. The fusion rule verification includes at least the verification of the funding agency and approval number, the result time window verification, and the patentee consistency verification. The result time window verification specifically adopts a differentiated domain semantic time window expansion model: September 1st of the previous calendar year of the project execution period is the total starting point of the time window, and the next calendar year of the project execution period is the total ending point of the time window. For paper-type results, the starting point of the time window is adjusted to January 1st of the execution period start year. For patent-type results, the patent application date or authorization announcement date is used as the time determination field. S6. Output the full retrieval process and verification conclusions in a structured manner, and append the information of the verified result items to the verification cache record.

2. The intelligent review method for scientific research achievements oriented towards the progress and completion report of funded projects as described in claim 1, characterized in that, In step S2, the positioning result section specifically includes: A constraint-driven semantic perception and localization method for weakly structured long text outputs is adopted. The expected order of occurrence of output chapter names is preset, and all candidate positions of each chapter name are retrieved in the full text. The candidate set is filtered by the constraint that the index values ​​of each candidate position are strictly monotonically increasing, and the set of chapters that satisfy the desired order and whose position values ​​are continuously increasing is retained. For frequently occurring keywords, the adjacent suffix pattern is detected through a contextual sliding window, and candidate positions that match the exclusion pattern are filtered during the retrieval stage.

3. The intelligent review method for scientific research achievements oriented towards the progress and completion report of funded projects, as described in claim 1, is characterized in that... In step S2, during the structured information extraction, for English results entries, a result title boundary inference algorithm oriented towards academic literature citation formats is specifically adopted: Filter out Chinese characters in the entries to select a set of valid English tokens; The trailing anchor tokens are determined in four priority levels: the first priority is the journal or conference name token containing compound academic identifier phrases; the second priority is the token containing conference identifiers and also containing year numbers within a preset year range; the third priority is the year token consisting of four digits; and the fourth priority is the second to last item in the valid token set as the fallback anchor. Using the identified anchor token as the front boundary delimiter, the first paragraph of the article is truncated and the last semicolon-separated unit is extracted as the first title candidate; at the same time, the token with the most words, after excluding tokens already identified as journal names or conference names, is used as the second fallback candidate. Both will be combined to form a set of multiple candidate titles.

4. The intelligent review method for scientific research achievements oriented towards the progress and completion report of funded projects as described in claim 1, characterized in that, In step S4, dynamically determining the calling order specifically involves: A two-dimensional routing matrix is ​​constructed with output type and language attribute as dimensions; the determination of the language attribute of the output adopts a two-stage mechanism: first, the proportion of Chinese characters is counted, and when it exceeds the preset threshold, it is determined to be Chinese; otherwise, a statistical language model is used for classification. For journal articles or conference papers in English, English databases are used first, followed by Chinese databases; for journal articles or conference papers in Chinese, Chinese databases are used first. For patent-related achievements, the patent database is specifically accessed, and a precise matching strategy based on the patent application number is used for matching.

5. The intelligent review method for scientific research achievements oriented towards the progress and completion report of funded projects as described in claim 1, characterized in that, In step S4, the semantic similarity algorithm specifically includes: Set the first threshold. Second threshold and author auxiliary threshold ,in ; For each candidate record in the search results list, calculate the semantic similarity between its result name and the target result name based on the token set. And the similarity between the author list and the target author. ; when When it is directly judged as a hit; when and When the time is right, it is considered a hit; Take all hit candidates The largest one is used as the final matching result when The search will terminate immediately if the score is full.

6. The intelligent review method for scientific research achievements oriented towards the progress and completion report of funded projects, as described in claim 1, is characterized in that... In step S5, the verification of the funding agency and approval number further includes: When the funding category of a project belongs to a type with a sub-project hierarchical structure, the prefix field of the parent project approval number is taken and matched with the sub-project approval number by truncation and inclusion.

7. The intelligent review method for scientific research achievements oriented towards the progress and completion report of funded projects as described in claim 1, characterized in that, In step S5, the patent holder consistency verification specifically employs a keyword inclusion matching strategy: The verification is considered successful when the patentee field returned by the search contains any of the following forms: the full name, abbreviation, English name of the affiliated unit, or the name of its subordinate second-level unit.

8. The intelligent review method for scientific research achievements oriented towards the progress and completion report of funded projects, as described in claim 1, is characterized in that... Also includes: Based on the document-level verification pass rate, the processing results of each report are managed hierarchically, and all related products generated by the same report are copied to the corresponding level directory to form a visual hierarchical archiving structure.

9. A research achievement intelligent review system based on any one of claims 1-8, which is a method for intelligent review of research achievements related to the progress and completion reports of funded projects, characterized in that, include: The results extraction module is used to receive and parse the progress report or completion report of the funded project, perform text standardization and cleaning, extract project identification information from the first page of the report, locate the results chapter in the cleaned report text and divide it into independent results items, and perform information structure extraction on each results item to obtain results information that includes at least the results name, results category, author and publication year. The verification cache management module is used to query verification cache records based on a composite cache key consisting of the project approval number, achievement category, and achievement name of the current achievement item. If a match is found, it is directly marked as verified and subsequent searches are skipped. If no match is found, the multi-source search module is triggered. The module is also used to append the information of verified achievement items to the verification cache record. The multi-source retrieval module is used to dynamically determine and call at least one external academic database for retrieval based on the category and language attributes of the results, and to use a semantic similarity algorithm based on token sets to match the retrieval results with the results information to obtain candidate matching results. The consistency verification module is used to obtain the metadata of candidate matching results, map it to a unified verification field, and perform fusion rule verification with the project identification information and result information extracted from the report homepage to generate item-level verification conclusions; the fusion rule verification includes at least the verification of funding agency and approval number, result time window verification, and patent holder consistency verification. The results output module is used to output the full retrieval process and verification conclusions in a structured manner.

Images